Thyroid disease has become the second largest disease in the field of endocrinology. SPECT imaging is especially important for the clinical diagnosis of thyroid disease. In order to reduce the workload of clinicians, reduce the rate of misdiagnosis, and improve the accuracy of deep learning algorithm in recognizing the features of cross images in nuclear medical image-assisted diagnosis. The deep convolution generated countermeasures network (DCGAN) and high resolution generated countermeasures network (SRGAN) were used to generate images and improve the resolution to make up for the deficiency of training data.At the same time, the residual block output information is added to the cross-feature image information, and the learning of the cross-feature is added on the basis of retaining the learned image features, so as to improve the model. For cross-image features, a cross-training set is proposed to retrain the improved ResNet neural network model that has been trained with a single feature image.The experimental results show that after 100 rounds of iteration, the verification accuracy of the improved residual neural network model trained by the cross-training set is as high as 96.33%, and the verification loss is reduced to 0.1187 and tends to be stable.The recall rate, precision rate, specificity and F1 score were all above 93.8% in the recognition results. Finally, the improved neural network model and the new training method showed higher typical symptom recognition rate for thyroid SPECT imaging than other methods based on convolutional neural network (CNN).